Apertus 1.5 Release Notes

Project Apertus released Apertus 1.5 on July 24, 2026, publishing updated 8B and 70B weights on Hugging Face under an Apache 2.0 license. The release extends the first-generation Apertus 1.0 models with image and audio understanding, an optional thinking mode, a context window four times longer than before, stronger instruction-following, better tool use, and closer adherence to the project's stated values. Full details appear in the project Technical Report, cited as arXiv:2610.15542.
The models were built by continued pretraining rather than training from scratch. Apertus 1.5 adds 4 trillion tokens of text and multimodal data to the 8B model and 2 trillion to the 70B model, bringing their totals to 19T and 17T tokens respectively. As with earlier Apertus releases, the project describes the models as fully open, covering weights, data, values, and training details, and it publishes permissively licensed data transparency records.
Multimodality is handled through the tokenizer rather than a separate vision encoder. Images and audio are converted into discrete tokens and placed in the same sequence as text, a design the project calls early fusion. Images pass through the Emu3.5 tokenizer, where a 512x512 image becomes 1,024 tokens, and audio through WavTokenizer at 40 tokens per second. Because this only extends the vocabulary, the underlying architecture is unchanged and training runs at text-only speed. The models accept images and audio but generate text only, and spoken-language processing is described as experimental.
For developers, the most concrete changes are the context window and the reasoning switch. The window grows to 262,144 tokens, four times the Apertus 1.0 limit, which in practice allows whole repositories, long documents or extended conversation histories in a single prompt, at the cost of more memory and compute per request. Thinking mode makes the model reason over the input before answering, which the project links to better results on reasoning problems. Instruction-following and tool integration were also trained for directly, which should matter to anyone wiring the models into agents or API-calling pipelines.
Training ran in four pretraining stages beginning from Apertus 1.0: an unpaired stage mixing text with standalone images and audio; a paired stage using image-caption and audio-transcript pairs to align the new modalities with language; a cooldown stage with higher-quality data weighted toward math and code; and a long-context stage that grows the window step by step from 8k to 262k tokens. More than half of all training tokens remain text-only, which the project says preserves Apertus 1.0's language abilities. Text data was refreshed from sources including FineWeb2-HQ+, FinePDFs and Nemotron-CC, and all new data, multimodal included, meets the project's stated bar of being publicly available, permissively licensed, and compliant with robots.txt opt-outs retroactively.
Post-training proceeds through five stages: supervised fine-tuning, long-context SFT, reinforcement learning, direct preference optimization, and charter alignment via SDPO. SFT used 3.7 million text samples plus image and speech data; reasoning traces were added only to prompts the model could not already solve without them, and tool-use data was restricted to execution-verified trajectories. The text SFT mix has been released on Hugging Face. Long-context SFT at the full 262k window is used to restore long-context ability that short-context SFT erodes, a limitation many practitioners will recognize. Reinforcement learning used verifiable rewards across roughly 350,000 math, code, instruction-following, multilingual and vision prompts with thinking both on and off. DPO used judge-scored preference pairs to improve chat quality, tool use and safety, and the preference pool and training pairs are public. SDPO aligns the model with the Apertus Charter through judge feedback and self-distillation. This suggests the release is aimed at teams that need auditability as much as raw capability, since each stage's data and checkpoints can be inspected.
Evaluation scope was expanded alongside the capabilities. Multimodal testing runs through an MLLM-eval-suite pipeline that merges the lmms-eval and VLMEvalKit frameworks, and new text benchmarks cover translation, multilingual ability, safety and tool use. The project reports that Apertus 1.5 still trails frontier open-weight models overall, but improved significantly in coding and instruction following, with the 8B model outperforming similarly sized open-weight models on instruction-following and multilingual tasks. Thinking mode further lifts performance on reasoning-heavy benchmarks such as AIME. Multimodal performance is described as closer to Gemma 4 12B and other current open-weight models, competitive on spatial understanding among open-weight models and on document and chart understanding among open-source models. On Swiss Legal tasks the release is reported on par with leading open-weight models, and on the Last Translation Benchmark the 8B model is said to match GPT-OSS-20B and outperform Claude Haiku 4.5 and Google Translate.
A notable structural choice is the evaluation reporting. Apertus publishes results with the EvalEval Coalition in the Every Eval Ever schema, recording not just scores but who ran the evaluation, how the model was accessed, generation settings and how each metric should be interpreted. The project notes EvalEval has mapped Evaluation Cards to AI governance requirements including Article 55 of the EU AI Act and the GPAI Code of Practice, and found the EEE schema can represent 56 of the 60 reporting elements those frameworks call for. Because the text evaluations run on a public fork of lm-evaluation-harness that EEE supports with a converter, the numbers can be rerun and compared under identical settings. For teams that have to justify model choices to compliance reviewers, this kind of machine-readable record is more useful than a leaderboard position.
Reproducibility extends to intermediate artifacts. Base weights, SFT, RLVR and DPO checkpoints are hosted in an Apertus 1.5 Hugging Face collection, with training data mixes in a companion datasets collection, so researchers can trace how each stage changes behavior. The report documents ablations on post-training data, reporting that multimodal data mixed with text during SFT improves results across all modalities, and that benefits from multimodal reasoning traces survive later text-only reinforcement learning: starting from an SFT checkpoint trained with multimodal traces, RL improves mathematics more than from a text-only checkpoint, and the reasoning advantage persists.
On availability, Proton Lumo and Infomaniak Euria have integrated Apertus 1.5, and API access has been announced by providers including Swisscom, Infomaniak and Phoeniqs, with others in the project's General Availability guide. Technical instructions live in the model cards and a Quickstart guide. The project also marked five million Hugging Face downloads and announced a roadmap for Apertus 2.0 at the AI+X Summit in Zurich: a sparse/MoE architecture of roughly 600B total parameters, an extended training schedule, and priorities around agents, mathematics and coding.
Why it matters: teams that want an open-weights model with long context, multimodal input and documented provenance now have a concrete Apache 2.0 option in 8B and 70B sizes, plus API routes through several European providers. The thinking-mode switch and stronger tool use make it plausible for agent and retrieval workloads, though output is text-only and audio handling is experimental. The checkpoint and evaluation releases lower the cost of verifying claims internally, which is likely to matter most to regulated or public-sector deployments; whether the 8B and 70B models are strong enough to displace larger alternatives is something each team will have to test on its own tasks.
Based on reporting from the original publisher. Visit the source for full context and later updates.